evalscope 0.17.1__py3-none-any.whl → 1.0.1__py3-none-any.whl

evalscope/api/benchmark/adapters/default_data_adapter.py

@@ -0,0 +1,684 @@
+import os
+from collections import defaultdict
+from functools import partial
+from overrides import override
+from typing import Any, Callable, Dict, List, Optional, Tuple, Type
+
+from evalscope.api.dataset import DataLoader, Dataset, DatasetDict, LocalDataLoader, RemoteDataLoader, Sample
+from evalscope.api.evaluator import TaskState
+from evalscope.api.messages import ChatMessage, ChatMessageSystem, ChatMessageUser
+from evalscope.api.metric import AggScore, SampleScore, Score
+from evalscope.api.model import Model, ModelOutput
+from evalscope.api.registry import get_aggregation, get_metric
+from evalscope.constants import HubType, JudgeStrategy
+from evalscope.report import Report, ReportGenerator
+from evalscope.utils import get_logger
+from ..benchmark import DataAdapter
+
+logger = get_logger()
+
+
+class DefaultDataAdapter(DataAdapter):
+    """
+    Default Data Adapter for the benchmark evaluation system.
+
+    This class serves as the base implementation for data adapters that handle:
+    - Dataset loading and preprocessing
+    - Model inference execution
+    - Metric calculation and aggregation
+    - Report generation
+
+    The adapter follows a pipeline architecture with hooks that can be overridden
+    in subclasses to customize behavior for specific benchmarks or evaluation tasks.
+
+    Key responsibilities:
+    1. Load datasets with optional few-shot examples
+    2. Process samples and format prompts
+    3. Execute model inference with proper state management
+    4. Calculate evaluation metrics and aggregate results
+    5. Generate comprehensive evaluation reports
+
+    This class can be extended to implement specific data loading and processing
+    logic for different benchmark datasets and evaluation scenarios.
+    """
+
+    # ####################
+    # DATA LOADING METHODS
+    # ####################
+
+    @override
+    def load_dataset(self) -> DatasetDict:
+        """
+        Load the complete dataset including test data and optional few-shot examples.
+
+        This method handles both local and remote dataset loading, processes samples
+        with appropriate prompt formatting, and prepares few-shot examples if needed.
+
+        Returns:
+            DatasetDict: A dictionary containing the loaded and processed datasets,
+                        organized by subset names.
+        """
+        # Load the dataset
+        self.test_dataset, self.fewshot_dataset = self.load()
+
+        # Process each sample's input by applying prompt templates and few-shot formatting
+        self._post_process_samples()
+
+        return self.test_dataset
+
+    def load(self) -> Tuple[DatasetDict, Optional[DatasetDict]]:
+        """Load the dataset from disk or remote source.
+
+        Returns:
+            Tuple[DatasetDict, Optional[DatasetDict]]: The test dataset and few-shot dataset.
+        """
+        if os.path.exists(self.dataset_id):
+            # Load dataset from local file system path
+            with self._temporary_attribute('dataset_hub', HubType.LOCAL):
+                return self.load_from_disk()
+        else:
+            # Load dataset from remote source (e.g., ModelScope, Huggingface)
+            return self.load_from_remote()
+
+    def load_from_remote(self):
+        """Load dataset from remote source and prepare few-shot examples if needed."""
+        test_dataset = None
+        fewshot_dataset = None
+        # Load dataset from remote source
+        test_load_func = partial(self.load_subset, data_loader=RemoteDataLoader)
+        test_dataset = self.load_subsets(test_load_func)
+
+        # Load few-shot examples if few-shot prompting is enabled
+        if self._should_load_fewshot():
+            fewshot_load_func = partial(self.load_fewshot_subset, data_loader=RemoteDataLoader)
+            fewshot_dataset = self.load_subsets(fewshot_load_func, is_fewshot=True)
+        return test_dataset, fewshot_dataset
+
+    def load_from_disk(self, use_local_loader: bool = False):
+        """
+        Load dataset from local disk path.
+
+        Args:
+            use_local_loader: If True, use local file loading; otherwise use remote loading
+                             for local ModelScope datasets.
+        """
+        test_dataset = None
+        fewshot_dataset = None
+        if use_local_loader:
+            # Use LocalDataLoader for actual local file loading
+            test_load_func = partial(self.load_subset, data_loader=LocalDataLoader)
+            test_dataset = self.load_subsets(test_load_func)
+
+            # Load few-shot examples if few-shot prompting is enabled
+            if self._should_load_fewshot():
+                fewshot_load_func = partial(self.load_fewshot_subset, data_loader=LocalDataLoader)
+                fewshot_dataset = self.load_subsets(fewshot_load_func, is_fewshot=True)
+            return test_dataset, fewshot_dataset
+        else:
+            # Fallback to remote loading for local ModelScope datasets
+            return self.load_from_remote()
+
+    def _should_load_fewshot(self) -> bool:
+        """Check if few-shot dataset should be loaded."""
+        return self.few_shot_num > 0 and self.train_split is not None
+
+    def _post_process_samples(self):
+        """Process all sample inputs with prompt formatting."""
+        for subset in self.test_dataset.keys():
+            for sample in self.test_dataset[subset]:
+                if isinstance(sample.input, str):
+                    sample.input = self.process_sample_str_input(sample, subset)
+
+    def process_sample_str_input(self, sample: Sample, subset: str) -> List[ChatMessage]:
+        """
+        Convert a sample's input string to a list of ChatMessage objects.
+
+        This method formats the sample input into a structured message format
+        suitable for model inference, including system prompts if configured.
+        """
+        input_text = self.process_sample_input(sample, subset=subset)
+        input_messages = [ChatMessageUser(content=input_text)]
+        if self.system_prompt:
+            input_messages.insert(0, ChatMessageSystem(content=self.system_prompt))
+        return input_messages
+
+    def process_sample_input(self, sample: Sample, subset: str) -> str:
+        """
+        Process a single sample's input by applying prompt templates and few-shot formatting.
+
+        This method handles the complete input preparation pipeline:
+        1. Retrieves few-shot examples if enabled
+        2. Formats few-shot examples into demonstration text
+        3. Applies appropriate prompt template (with or without few-shot context)
+
+        Args:
+            sample (Sample): The sample to process
+            subset (str): The subset name this sample belongs to
+
+        Returns:
+            str: The formatted input text ready for model inference
+        """
+        if self.few_shot_num > 0:
+            if self.fewshot_dataset is not None:
+                # Retrieve few-shot examples for the current subset
+                few_shot_samples = self.fewshot_dataset.get(subset)
+                if few_shot_samples is None:
+                    # Fallback: use the first available subset if current subset not found
+                    first_key = next(iter(self.fewshot_dataset))
+                    few_shot_samples = self.fewshot_dataset[first_key]
+                # Select fewshot samples
+                assert len(few_shot_samples) >= self.few_shot_num, (
+                    f"""The dataset only have ({len(few_shot_samples)}) few-shot samples, but requested ({self.few_shot_num}) fewshot samples, please reduce 'few_shot_num'."""  # noqa: E501
+                )
+                # Convert few-shot samples to demonstration string
+                few_shot = '\n\n'.join([self.sample_to_fewshot(sample) for sample in few_shot_samples])
+            else:
+                # Build few-shot examples inside the format method
+                few_shot = ''
+            # Format the input text with few-shot examples and main prompt
+            input_text = self.format_fewshot_template(fewshot=few_shot, sample=sample)
+        else:
+            # No few-shot examples: use the prompt template directly
+            input_text = self.format_prompt_template(sample=sample)
+        return input_text
+
+    def load_subsets(self, load_func: Callable[[str], Dataset], is_fewshot=False) -> DatasetDict:
+        """
+        Load multiple subsets of the dataset using the provided loading function.
+
+        This method handles two loading strategies:
+        1. Reformat mode: Load only the default subset and reformat it
+        2. Multi-subset mode: Load all subsets specified in subset_list
+
+        Args:
+            load_func (Callable[[str], Dataset]): Function to load individual subsets
+
+        Returns:
+            DatasetDict: Dictionary containing all loaded subsets
+        """
+        if self.reformat_subset:
+            # Load only the default subset
+            subset_data = load_func(self.default_subset)
+            # Reformat the subset to create multiple subsets based on sample keys
+            # NOTE: subset_list and limit is applied here if specified
+            limit = self.few_shot_num if is_fewshot else self.limit
+            repeats = 1 if is_fewshot else self.repeats
+            dataset_dict = DatasetDict.from_dataset(
+                dataset=subset_data, subset_list=self.subset_list, limit=limit, repeats=repeats
+            )
+        else:
+            # Load all specified subsets into separate entries
+            subset_dict = defaultdict()
+            for subset in self.subset_list:
+                # Set current subset, since same benchmark need to differentiate
+                with self._temporary_attribute('current_subset_name', subset):
+                    subset_data = load_func(subset)
+                    subset_dict[subset] = subset_data
+            dataset_dict = DatasetDict(subset_dict)
+        return dataset_dict
+
+    def load_subset(self, subset: str, data_loader: Type[DataLoader]) -> Dataset:
+        """
+        Load a specific subset of the dataset for evaluation.
+
+        Args:
+            subset (str): The subset identifier to load
+            data_loader (Type[DataLoader]): The data loader class to use for loading
+
+        Returns:
+            Dataset: The loaded dataset subset with processed samples
+        """
+        # Determine the split and subset names based on configuration
+        split = subset if self.split_as_subset else self.eval_split
+        subset_name = self.default_subset if self.split_as_subset else subset
+
+        # Create and configure the remote data loader
+        loader = data_loader(
+            data_id_or_path=self.dataset_id,
+            split=split,
+            subset=subset_name,
+            sample_fields=self.record_to_sample,  # Custom sample conversion function
+            filter_func=self.sample_filter,
+            limit=self.limit if not self.reformat_subset else None,  # Limit number of samples if specified
+            repeats=self.repeats,  # Number of repetitions for each sample
+            shuffle=self.shuffle,  # Shuffle dataset if enabled
+            shuffle_choices=self.shuffle_choices,  # Shuffle choices if requested
+            data_source=self.dataset_hub,  # Data source configuration
+        )
+        dataset = loader.load()
+        return dataset
+
+    def load_fewshot_subset(self, subset: str, data_loader: Type[DataLoader]) -> Dataset:
+        """
+        Load a subset specifically for few-shot examples.
+
+        Args:
+            subset (str): The subset identifier to load few-shot examples from
+            data_loader (Type[DataLoader]): The data loader class to use for loading
+
+        Returns:
+            Dataset: The loaded few-shot dataset with demonstration examples
+        """
+        # Use training split for few-shot examples
+        split = subset if self.split_as_subset else self.train_split
+        subset_name = self.default_subset if self.split_as_subset else subset
+
+        # Create loader specifically configured for few-shot sampling
+        loader = data_loader(
+            data_id_or_path=self.dataset_id,
+            split=split,
+            subset=subset_name,
+            sample_fields=self.record_to_sample,
+            filter_func=self.sample_filter,  # Apply sample filtering if defined
+            limit=self.few_shot_num
+            if not self.reformat_subset else None,  # Limit to specified number of few-shot examples
+            shuffle=self.few_shot_random,  # Randomize selection if enabled
+            shuffle_choices=self.shuffle_choices,  # Shuffle choices if requested
+            data_source=self.dataset_hub,
+        )
+        dataset = loader.load()
+        return dataset
+
+    def sample_filter(self, sample: Sample) -> bool:
+        """
+        Apply filtering to a dataset, only samples matching the predicate will be included.
+
+        Args:
+            sample (Sample): The sample to filter
+
+        Returns:
+            bool: True if the sample passes the filter, False otherwise
+        """
+        return True  # Default implementation allows all samples
+
+    def record_to_sample(self, record: Dict[str, Any]) -> Sample:
+        """
+        Convert a raw data record to a Sample object.
+
+        This method must be implemented in subclasses to handle dataset-specific
+        field mapping and data processing logic.
+
+        Args:
+            record (Dict[str, Any]): Raw data record from the dataset
+
+        Returns:
+            Sample: Processed sample object ready for evaluation
+        """
+        raise NotImplementedError('This method should be implemented in subclasses')
+
+    def sample_to_fewshot(self, sample: Sample) -> str:
+        """
+        Convert a Sample object to a formatted few-shot demonstration string.
+
+        This method must be implemented in subclasses to define how samples
+        are formatted as examples in few-shot prompts.
+
+        Args:
+            sample (Sample): The sample to convert to a few-shot example
+
+        Returns:
+            str: Formatted few-shot demonstration string
+        """
+        raise NotImplementedError('This method should be implemented in subclasses')
+
+    def format_prompt_template(self, sample: Sample) -> str:
+        """
+        Format the basic prompt template with the sample data.
+
+        This method applies the prompt template to format the input text
+        for models when no few-shot examples are used.
+
+        Args:
+            sample (Sample): The sample object containing the prompt data
+
+        Returns:
+            str: The formatted prompt ready for model input
+        """
+        return self.prompt_template.format(question=sample.input)
+
+    def format_fewshot_template(self, fewshot: str, sample: Sample) -> str:
+        """
+        Format the few-shot template with demonstrations and the main prompt.
+
+        This method combines few-shot examples with the main prompt using
+        the configured few-shot template.
+
+        Args:
+            fewshot (str): The formatted few-shot demonstration examples
+            sample (Sample): The sample object containing the prompt data
+
+        Returns:
+            str: The complete formatted input with few-shot context
+        """
+        return self.few_shot_prompt_template.format(fewshot=fewshot, question=sample.input)
+
+    # #################
+    # INFERENCE METHODS
+    # #################
+
+    def _on_inference_start(self, model: Model, sample: Sample) -> None:
+        """
+        Hook method called before inference starts.
+
+        This method can be overridden in subclasses to implement custom
+        preparation logic before model inference (e.g., model configuration,
+        sample preprocessing, state initialization).
+
+        Args:
+            model (Model): The model that will perform inference
+            sample (Sample): The sample to be processed
+        """
+        pass
+
+    def _on_inference(self, model: Model, sample: Sample) -> ModelOutput:
+        """
+        Hook method called during the actual inference process.
+
+        This method executes the model inference and can be overridden
+        to implement custom inference logic or model interaction patterns.
+
+        Args:
+            model (Model): The model to use for inference
+            sample (Sample): The sample to process
+
+        Returns:
+            ModelOutput: The raw output from the model
+        """
+        # Execute model inference with the processed input and any tools
+        model_output = model.generate(input=sample.input, tools=sample.tools)
+        return model_output
+
+    def _on_inference_end(
+        self, model: Model, sample: Sample, model_output: ModelOutput, output_dir: str, **kwargs
+    ) -> TaskState:
+        """
+        Hook method called after inference completes.
+
+        This method processes the model output and creates a TaskState object
+        that encapsulates all information about the completed inference task.
+        You can save the model output to the specified output directory.
+
+        Args:
+            model (Model): The model that performed inference
+            sample (Sample): The processed sample
+            model_output (ModelOutput): The raw model output
+            output_dir (str): The directory where the model output was saved
+
+        Returns:
+            TaskState: Complete state object for the inference task
+        """
+        return TaskState(
+            model=model.name,
+            sample=sample,
+            messages=[model_output.message],
+            output=model_output,
+            completed=True,
+        )
+
+    @override
+    def run_inference(self, model: Model, sample: Sample, output_dir: str, **kwargs) -> TaskState:
+        """
+        Execute the complete inference pipeline for a single sample.
+
+        This method orchestrates the full inference process using the hook methods:
+        1. Pre-inference preparation
+        2. Model inference execution
+        3. Post-inference processing and state creation
+
+        Args:
+            model (Model): The model to use for inference
+            sample (Sample): The sample to process
+            output_dir (str): The directory to store the generated files
+
+        Returns:
+            TaskState: Complete state object containing inference results
+        """
+        self._on_inference_start(model, sample)
+        model_output = self._on_inference(model, sample)
+        task_state = self._on_inference_end(model, sample, model_output, output_dir, **kwargs)
+
+        return task_state
+
+    # ##########################
+    # METRIC CALCULATION METHODS
+    # ##########################
+
+    def filter_prediction(self, prediction: str, task_state: TaskState) -> str:
+        """
+        Filter and prepare the model prediction for metric calculation.
+
+        This method applies configured filters and custom answer extraction
+        to clean and prepare the raw model output for evaluation.
+
+        Args:
+            prediction (str): The raw model prediction
+            task_state (TaskState): The complete task state for context
+
+        Returns:
+            str: The filtered and extracted prediction ready for evaluation
+        """
+        if self.filter_ensemble is not None:
+            # Apply configured filters to clean the prediction
+            prediction = self.filter_ensemble(prediction)
+
+        # Apply custom answer extraction logic
+        extracted_prediction = self.extract_answer(prediction, task_state)
+        return extracted_prediction
+
+    def extract_answer(self, prediction: str, task_state: TaskState) -> str:
+        """
+        Hook method for custom answer extraction from model predictions.
+
+        This method can be overridden in subclasses to implement specific
+        logic for extracting the final answer from complex model outputs.
+
+        Args:
+            prediction (str): The model prediction to extract from
+            task_state (TaskState): The task state for additional context
+
+        Returns:
+            str: The extracted answer
+        """
+        return prediction
+
+    def match_score(
+        self, original_prediction: str, filtered_prediction: str, reference: str, task_state: TaskState
+    ) -> Score:
+        """
+        Calculate evaluation scores by comparing prediction with reference.
+
+        This method computes scores using all configured metrics and creates
+        a comprehensive Score object with detailed evaluation results.
+
+        Args:
+            original_prediction (str): The original, unfiltered model prediction
+            filtered_prediction (str): The filtered and processed prediction
+            reference (str): The ground truth reference answer
+            task_state (TaskState): The complete task state for context
+
+        Returns:
+            Score: Object containing all calculated metric scores and metadata
+        """
+        # Initialize the score object with prediction details
+        score = Score(
+            extracted_prediction=filtered_prediction,
+            prediction=original_prediction,
+        )
+
+        # Calculate scores for each configured metric
+        for metric in self.metric_list:
+            try:
+                if isinstance(metric, str):
+                    metric_name = metric
+                    metric_scorer = get_metric(metric)  # Get metric implementation from registry
+                    metric_func = metric_scorer()  # Instantiate the metric scorer
+                elif isinstance(metric, dict):
+                    metric_name = list(metric.keys())[0]
+                    metric_cls = get_metric(metric_name)
+                    metric_func = metric_cls(**metric[metric_name])  # Initialize with parameters
+                metric_score = metric_func(
+                    prediction=filtered_prediction,
+                    reference=reference,
+                )
+                score.value[metric_name] = metric_score
+            except Exception as e:
+                logger.error(f'Error calculating metric {metric}: {e}')
+                score.value[metric_name] = 0
+                score.metadata[metric_name] = f'error: {str(e)}'
+
+        return score
+
+    @override
+    def calculate_metrics(self, task_state: TaskState) -> SampleScore:
+        """
+        Calculate comprehensive evaluation metrics for a completed task.
+
+        This method processes the task state to extract predictions, applies
+        filtering and answer extraction, calculates all configured metrics,
+        and packages the results into a SampleScore object.
+
+        Args:
+            task_state (TaskState): The completed task state to evaluate
+
+        Returns:
+            SampleScore: Complete scoring results for the sample
+
+        Raises:
+            AssertionError: If the task state is not marked as completed
+        """
+        assert task_state.completed, \
+            'TaskState must be completed before calculating metrics.'
+
+        # Extract the raw prediction from the model output
+        prediction = task_state.output.completion
+
+        # Apply filtering and answer extraction
+        filtered_prediction = self.filter_prediction(prediction, task_state)
+
+        if self.judge_strategy == JudgeStrategy.LLM_RECALL:
+            # Step 1: Calculate standard metric scores (rule-based)
+            rule_based_score = self.match_score(
+                original_prediction=prediction,
+                filtered_prediction=filtered_prediction,
+                reference=task_state.target,
+                task_state=task_state
+            )
+
+            # Step 2: Apply LLM judge if enabled and get final score
+            final_score = self.maybe_llm_match_score(
+                original_prediction=prediction,
+                filtered_prediction=filtered_prediction,
+                reference=task_state.target,
+                task_state=task_state,
+                rule_based_score=rule_based_score
+            )
+        else:
+            if self.use_llm_judge:
+                # Use LLM judge to compute the match score directly
+                final_score = self.llm_match_score(
+                    original_prediction=prediction,
+                    filtered_prediction=filtered_prediction,
+                    reference=task_state.target,
+                    task_state=task_state
+                )
+            else:
+                # Use standard match score calculation without LLM judge
+                final_score = self.match_score(
+                    original_prediction=prediction,
+                    filtered_prediction=filtered_prediction,
+                    reference=task_state.target,
+                    task_state=task_state
+                )
+
+        # Package the results into a sample score object
+        sample_score = SampleScore(
+            score=final_score,
+            sample_id=task_state.sample_id,
+            group_id=task_state.group_id,
+            sample_metadata=task_state.metadata,
+        )
+
+        return sample_score
+
+    @override
+    def aggregate_scores(self, sample_scores: List[SampleScore]) -> List[AggScore]:
+        """
+        Aggregate individual sample scores into summary statistics.
+
+        This method uses the configured aggregation method to compute
+        summary statistics (e.g., mean, median, percentiles) across
+        all sample scores for comprehensive evaluation results.
+
+        Args:
+            sample_scores (List[SampleScore]): Individual scores for all samples
+
+        Returns:
+            List[AggScore]: Aggregated scores and statistics
+        """
+        # Get the configured aggregation implementation
+        aggregate_cls = get_aggregation(self.aggregation)
+        aggregator = aggregate_cls()
+
+        # Compute aggregated scores
+        agg_scores = aggregator(sample_scores)
+
+        return agg_scores
+
+    # #########################
+    # REPORT GENERATION METHODS
+    # #########################
+
+    def _on_generate_report_end(self, report: Report, output_dir: str, **kwargs) -> None:
+        """
+        Hook method called after generating the evaluation report.
+
+        This method can be overridden in subclasses to implement custom
+        post-processing of the generated report (e.g., additional formatting,
+        custom visualizations, external integrations).
+
+        Args:
+            report (Report): The generated evaluation report
+            output_dir (str): Directory where the report should be saved
+        """
+        pass
+
+    def _on_generate_report(
+        self, scores: Dict[str, List[AggScore]], model_name: str, add_aggregation_name: bool = True
+    ) -> Report:
+        """
+        Hook method called during report generation.
+
+        This method creates the evaluation report using the configured
+        report generator and can be overridden to implement custom
+        report generation logic.
+
+        Args:
+            scores (Dict[str, List[AggScore]]): Aggregated scores organized by subset
+            model_name (str): Name of the evaluated model
+
+        Returns:
+            Report: The generated evaluation report
+        """
+        return ReportGenerator.generate_report(
+            score_dict=scores, model_name=model_name, data_adapter=self, add_aggregation_name=add_aggregation_name
+        )
+
+    @override
+    def generate_report(self, scores: Dict[str, List[AggScore]], model_name: str, output_dir: str, **kwargs) -> Report:
+        """
+        Generate a comprehensive evaluation report from aggregated scores.
+
+        This method orchestrates the complete report generation process:
+        1. Creates the report using configured generators
+        2. Applies any post-processing through hook methods
+
+        Args:
+            scores (Dict[str, List[AggScore]]): Aggregated scores by subset name
+            model_name (str): Name of the model being evaluated
+
+        Returns:
+            Report: Complete evaluation report with results and analysis
+        """
+        report = self._on_generate_report(scores, model_name=model_name)
+        self._on_generate_report_end(report, output_dir, **kwargs)
+        return report